dataset card

README.md

@@ -1,120 +1,5 @@
 # handTrackingSample
 
-Egocentric footage of people working with their hands, cut to the sections a human picked as
-hand-tracked, with synchronised IMU and hand pose on **every frame**.
-
-**120 episodes · 8.5 hours · 44 recordings · 27 cameras · 1920x1200 · stereo**
-
----
-
-> **19 Aug 2026 — full re-release.** Every episode here was re-rendered from raw: no lens
-> rectification (the earlier release had a visible circular seam on cameras whose undistortion map
-> folds inside the frame), a stricter face-blur setting (see *Anonymisation*), no orientation flip
-> guessing, and WiLoR hand pose on every frame rather than every sixth. The earlier 149-episode
-> set has been withdrawn in full; this set replaces it.
-
-
-## What each episode contains
-
-```
-<run>/<seg>/
-  left.mp4      1920x1200 H.265, white-balanced, faces pixelated (native lens geometry)
-  right.mp4     the other eye of the stereo pair, same treatment
-  frames.csv    frame_idx, ts_us      <- the authoritative video clock
-  imu.csv       ts_us, ax/ay/az (m/s2), gx/gy/gz (deg/s)
-  hands.jsonl   per-frame hand boxes, 21-joint 2D/3D pose and MANO parameters
-  labels.json   scene, task, job, environment, zone provenance
-  _stat.json    processing record, including the exact anonymisation setting
-```
-
-## How the sections were chosen
-
-A person reviewed the recordings and drew the hand-tracked ranges by hand. Those ranges -- not a
-detector -- define the episodes here. An independent hand detector agrees with them: every curated
-zone has hand detections in at least 52.9% of its frames, median 92.3%.
-
----
-
-## Read this before you align anything
-
-**Use `frames.csv`, never the container frame rate.** The mp4 carries a synthetic constant-rate
-PTS. The true exposure clock differs from it by up to **135 ms across an episode** -- about four
-frames. Anyone computing frame time as `index / fps` will drift out of sync with the IMU by the
-end. `frames.csv` carries the real per-frame timestamp, on the same clock as `imu.csv`.
-
-**`dt` is not uniform.** Recordings contain genuine dropped-frame gaps, up to about **0.5 s**.
-Video and IMU stay locked to each other across them -- the sample-per-frame ratio is constant -- but
-integrating gyro on an assumed fixed `dt` will accumulate error. Gaps are listed per episode in
-`_stat.json`.
-
-**`hands.jsonl` is keyed by `frame_idx`.** There is one row per decoded frame, but index by
-`frame_idx` rather than by position -- it is the same index `frames.csv` uses, and it is what ties
-a row to its timestamp.
-
----
-
-## Anonymisation -- known limitation
-
-Faces are detected with SCRFD-10G at a 960 letterbox, upright threshold **0.80**, unioned with a
-four-rotation pass at **0.80**, boxes over 10% of the frame additionally required to score 0.80,
-and pixelated with an oval mask held for 6 frames after the last detection.
-
-This is a deliberately **conservative** setting, chosen after the earlier release pixelated hands,
-cookware and packaging that the detector scored as faces. On a hand-labelled benchmark of 31 faces
-drawn from this same footage it catches **5 of 31 (16%)** with one false positive in 140 frames;
-the earlier release's setting caught 68% with eight. The faces it does catch are the frontal,
-close ones -- a bystander in profile or a face at an angle to the camera will usually score below
-0.80 and pass through unblurred.
-
-**16% is an upper bound on what was caught, not a floor on what leaked.** The benchmark can only
-contain faces some detector saw at all, so faces no model detected are absent from both the
-numerator and the denominator. **Identifiable faces remain in this data, more than in most
-published egocentric sets.** Do not use it for face recognition, re-identification, or any purpose
-requiring that individuals be unidentifiable. `_stat.json` records exactly what was applied, so
-a stronger pass can be layered on top with your own detector.
-
-## Lens geometry
-
-Frames are delivered in the camera's native geometry -- **not rectified**. The 116-degree optics
-are already close to rectilinear across most of the frame, and per-camera calibrations exist but
-roughly half of them produce an undistortion map that folds back on itself away from the
-calibration target, so applying them selectively would give an inconsistent dataset. Consistent
-raw geometry was judged the better trade. Per-camera intrinsics are available on request.
-
-## Hand pose
-
-21 3D joints per hand from **WiLoR** (ViT backbone + MANO mesh head), on **every frame**.
-MediaPipe's landmarker was evaluated against WiLoR and EgoHOS on this footage and was not good
-enough -- it is the same model family that came last of eight detectors on our face benchmark, and
-head-mounted video is its weak case.
-
-```json
-{"frame_idx": 120,
- "pose": [{"kp2d": [[x,y], ...21], "kp3d": [[x,y,z], ...21],
-           "is_right": true, "box": [x1,y1,x2,y2],
-           "mano": {"global_orient": [3], "hand_pose": [45], "betas": [10], "cam_t": [3]}}]}
-```
-
-- `kp2d` is in frame pixels; `box` is WiLoR's own hand detection, so detection and articulation
-  come from the same model and always agree.
-- `kp3d` is WiLoR's 3D joint prediction in its own metric frame (metres, hand-rooted), not camera
-  millimetres. `mano` carries the axis-angle global orientation, the 15 joint rotations, the shape
-  coefficients and the weak-perspective camera translation, so the full mesh can be rebuilt with a
-  standard MANO layer.
-- Left hands are predicted by mirroring, as in WiLoR; `kp3d.x`, `global_orient` and `hand_pose`
-  are already mirrored back into the un-flipped frame.
-- Frames with no detected hand carry `"pose": []`.
-- When a hand is inches from the lens its detection box can span half the frame and touch an
-  edge; the weak-perspective projection then places one or two of its `kp2d` joints well outside
-  the image. This affects roughly 1 hand in 100,000 -- clip or drop such joints as you see fit.
-
-Pose was computed with a batched re-implementation of WiLoR-mini's inference loop (crop-local
-anti-alias blur instead of a full-frame blur per hand). Against the stock pipeline on 102 hands it
-gives identical detections and a median 1.0 px / max 7.6 px difference in `kp2d`, which is the
-model's own sensitivity to sub-grey-level input rounding.
-
-**Held-object boxes are not in this release.** `hands.jsonl` carries hand detection and pose only.
-
 ## Scene and task spread
 
 120 episodes drawn from 44 recordings on 27 different cameras, across **13 top-level categories**:
@@ -138,16 +23,3 @@ model's own sensitivity to sub-grey-level input rounding.
 No category exceeds 8 of 44 recordings. Episodes were selected round-robin across cameras, so the
 set spans the full camera fleet rather than over-sampling whichever recordings happened to be
 longest.
-
-## How much of the footage carries a hand skeleton
-
-The share of frames that carry at least one hand varies a lot by task, and is a property of the
-work, not a failure of the pipeline: across the 120 episodes the median is **84%** of frames with at least one hand, the interquartile range 69–92%, and the lowest are the three pieces of `starcap52_0712-0010` (10%, 12%, 18%) -- an electrician working overhead, where the camera spends most of the episode on the ceiling and the hands enter from the top of frame, WiLoR's weak case.
-
-Check `hands.jsonl` before assuming a given episode is densely annotated.
-
-## Provenance
-
-Every episode records its source recording, the exact source frame range, the curated zone it came
-from, and the anonymisation setting used, in `_stat.json`. Cut and rendered from raw with
-`generateSample/run_sample.py`; pose with `generateSample/wilor_pose.py`.